FU Zhengyu

I'm a master student in Robotics, Systems and Control at ETH Zürich. Before that, I graduated from The Hong Kong University of Science and Technology with dual majors in Integrative System and Design and Computer Engineering.

From 2021 to 2022, I completed my bachelor thesis at the Robotic Systems Lab (RSL) at ETH Zurich on optimal control solvers for robotics, under the supervision of Prof. Marco Hutter and Dr. Farbod Farshidian .

In the summer of 2022, I was a Robotics Institute Summer Scholar (RISS) in the Robotic Exploration Lab at Carnegie Mellon University, supervised by Prof. Zachary Manchester.

I am a robotics enthusiast with broad interests in optimal control, control theory and applied optimization for agile robotics systems such as quadrupeds and quadrotors.

Github  /  LinkedIn

An open-source Visual-Inertial-Leg Odometry (VILO) state estimation solution, Cerberus, for legged robots that estimates position precisely on various terrains in real time using a set of standard sensors, including stereo cameras, IMU, joint encoders, and contact sensors.

arXiv

Existing scene graphs primarily capture spatial relationships but often overlook the functional roles of objects, limiting robots' ability to understand and interact with their environments effectively.

We introduce the Functional Scene Graph (FSG), which integrates both spatial and functional relationships among objects and their parts. The FSG is generated through a mapping pipeline powered by foundation models, while the uncertainty of ambiguous functional edges is quantified using probabilistic graphical models, with commonsense priors from visual-language models (VLMs). This uncertainty-aware FSG then guides the physical verification of ambiguous functional relationships in the real world, improving the robot's scene understanding and decision-making capabilities.

Demo for the functional scene graph

Shooting Basketballs with the Transformer Robot

We enable the Transformer Robot, a legged-wheel quadrupedal platform, to shoot basketballs from the free-throw line while crouched and self-correct shot errors. The shooting policy was trained in Isaac Lab using proximal policy optimization (PPO), and a neural event ODE-based correction module was implemented to model the shooting dynamics. By performing online gradient descent on policy parameters based on observed errors, the system iteratively refined shots to improve accuracy until successful scoring.

A numerical implementation of a parallelizable QP solver named Proportional-Integral Projected Gradient (PIPG) under the nonlinear MPC (NMPC) framework of OCS2 which verified the feasibility of boosting control frequency by parallel computing.

The implementation has been open sourced as the ocs2_slp package under the ocs2 repository. For more information, please see my Bachelor thesis.

Bachelor thesis

A novel primal-dual formulation of augmented Lagrangian named PDAL that can greatly mitigate the numerical issue associated with the ill-conditioned Hessian and speed up convergence.

Code

A concise LDL factorization routine dedicated for model predictive control. Loop-unrolling and code generation techniques are utilized to speed up factorization of problems with fixed-horizon.

Code

Many motion planning methods ignore temporal dimension of dynamic obstacles, frequently treating them as static at each timestamp. This approach often leads to becoming stuck in local minima or an inability to find solutions. Through the integration of Safe Interval Path Planning (SIPP) with Trajectory Optimization (TO), the planner can incorporate wait actions between active moves, exploring the temporal dimension of the problem and thereby avoiding certain local minima.

Video

I revised the parallelization scheme of the backward pass of different dynamical programming (DDP) in OCS2 which improved the performance be 18%. The pull requests (PRs) are merged into the main branch and the toolbox is available at OCS2.

A MPC-iLQR controller for differential wheeled robots that is capable of reasoning about dynamic obstacles and replanning online.

Video

Quadruped Walking Mechanism

A mechanical system consisting of four identical planar walking mechanisms and being propelled by friction with the ground.

Relative State Formation-based Multi-robot Objects Manipulation

A control algorithm that uses a group of identical robots to manipulate objects collectively.

IMU-based Motion Capture System

A motion capture system based solely on two IMUs attached to the forearm and elbow respectively.

Optimal Control (Graduate-level Course)

From this course, I learnt
unconstrained optimization,   Constrained Optimization,   Linear Quadratic Regulator(LQR),   Linear Quadratic Gaussian(LQG),   Convex MPC,   Dynamic Programming(DDP),  
Direct Collocation,   Sequential Quadratic Programming(SQP) and Optimization for Hybrid System

State Estimation, Planning and Control for Quadrotor

A comprehensive introduction to aerial robots. Topics include rigid-body dynamics, system modeling, control, trajectory planning, sensor fusion, and vision-based state estimation.

Thanks for the website design